On Consistency of Estimators in Simple Linear Regression 1

Transcription

1 On Consistency of Estimators in Simple Linear Regression 1 Anindya ROY and Thomas I. SEIDMAN Department of Mathematics and Statistics University of Maryland Baltimore, MD (anindya@math.umbc.edu) (seidman@math.umbc.edu) We derive a property of real sequences which can be used to provide a natural sufficient condition for the consistency of the least squares estimators of slope and intercept for a simple linear regression. KEY WORDS: Least squares estimators. 1 This has now appeared in Calcutta Statistical Assoc. Bulletin 53, pp , (2003).

2 I. Introduction Consider the simple linear regression model: y i = β 0 + β 1 x i + e i, i = 1, 2,..., n, (1) where {x i } is a fixed regressor sequence and e i i.i.d. normal (0, σ 2 ). Such examples are commonly encountered in practice, for example a linear trend model. It is clear that in applying (1) the regressor sequence cannot be completely arbitrary. Even if we were to have error-free data, it would obviously be necessary to have more than one x-value to determine both components of the parameter vector β = (β 0, β 1 ) for intercept and slope. This suggests as a natural condition for the statistical analysis that it would be desirable to avoid a situation in which {x i } converges to a limiting value. The point of this note is that this heuristic observation does, indeed, provide a simple sufficient condition for consistency: Theorem: If { x n } does not converge to a finite limit then one has consistency of the ordinary least squares estimators for both β 0 and β 1. Note that the variance of the least squares estimator, ˆβ, of the parameter vector β is σ 2 (X X) 1 where X is the n 2 matrix of columns of 1 and x i. Thus, one sufficient condition for the consistency of the whole vector is that (X X) 1 0, i.e., that the smallest eigenvalue of X X diverges to infinity. However, in most common text books (Fuller 1987, Ferguson 1996) the sufficient conditions are stated individually for 1

3 the intercept and slope estimates in terms of their respective variances, i.e., that V ar{ ˆβ 0 } V ar{ ˆβ 1 } = σ 2 (n 1 + s 2 n x 2 n) 0, = σ 2 s 2 n 0, (2) where x n = n 1 n i=1 x i and s 2 n = n i=1 (x i x n ) 2. In view of the sufficient conditions (2), our Theorem will be an immediate consequence of the fact, whose proof is our main contribution, that a sufficient condition for (2) is that the sequence {x i } is not convergent. II. Proof As noted in the Introduction, the following Lemma provides a natural sufficient condition for consistency of the least squares estimators, proving the Theorem. Lemma: Let {x i } 1 be an arbitrary sequence of real numbers which does not converge to a finite limit. Then, with { x n, s 2 n} as above, s 2 n x 2 n 0. Proof: Observe that s 2 n = n i=1 (x i ξ) 2 n(ξ x n ) 2 for arbitrary ξ so, taking ξ = x n+1, we have s 2 n+1 s 2 n = (x n+1 x n+1 ) 2 + n( x n+1 x n ) 2, (3) showing that {s 2 n} is an increasing sequence. We now consider several cases: 1. Suppose the sequence { x n } is unbounded. Fix n and let N be such that x N 2 x n for all N n. Then we will show that x k /s k 2/ n for k N. Fix k N and consider two subcases. 2

4 (a) Let x k 2 x n so x k x n x k /2. Then s 2 k = = k n x j x k 2 x j x k 2 k x j x n 2 + n x k x n 2 = s 2 n + n x k x n 2 n x k x n 2 n x 2 k /4 Therefore x k /s k 2/ n for k N. (b) Let x k < 2 x n x N. Then x k /s k < x N /s k. Because s k is a nondecreasing sequence we have x k /s k < x N /s N. The result then follows from the previous case. 2. Now suppose { x n } is bounded so it is now sufficient to show that s 2 n. We again consider two subcases: (a) Suppose { x n } does not converge to a limit and so has at least two convergent subsequences with distinct limit points: { x n1 } and { x n2 } converging to limit points a 1 and a 2, respectively, with a 1 a 2 so 0 < 2δ = a 1 a 2. Given any n we can find k 1 > n and then k 2 > k 1 such that x k1 { x n1 }, x k2 { x n2 }, and x k1 x k2 δ. Much as in the argument for part a in case 1, we now have s 2 k 1 s 2 k 2 + k 1 δ 2 s 2 n + nδ 2. Since {s 2 k } is nondecreasing, this shows that s 2 k. (b) Finally, if { x n } is convergent but {x n } is not, then [x n+1 x n ] 0, i.e., x n+1 x n δ infinitely often for some δ > 0, so (3) gives s 2 n. 3

5 III. Summary and Discussion We have shown that if the sequence of means, { x n } from a sequence of values for a fixed regressor in a simple linear regression does not converge to a finite limit, then the least squares estimates of the intercept and the slope are consistent. Of course in the case the sequence of means converges to a finite limit, we can still have consistent estimators as long as the sample sum of squares of deviations, s 2 n, diverges. The sequence s 2 n will not diverge if all but a finitely many values of the explanatory variable behave like a constant. It will be interesting to see if in the multiple regression case such characterizations are possible. IV. References Ferguson, T. S. (1996) A Course in Large Sample Theory. Chapman and Hall, London. Fuller, W. A. (1987) Measurement Error Models. Wiley, NY. 4